Kairos 1: A Model That Plays the USER, Not the Assistant (13 Wins of 17)

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Kairos 1: A Model That Plays the USER, Not the Assistant (13 Wins of 17) · Channel: Prompt Engineer 48 · 03:08 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI models

Summary

Kairos 1 is presented as a model designed to simulate the human user rather than act as an assistant. On the company’s benchmarks, it scores 86.11 on the user simulation index and leads in 13 of 17 comparisons, including large gaps on some behavioral tests. The video notes that the results come from an early checkpoint, some comparisons are close, and the figures are presented by Diachronic rather than independently verified. The potential use suggested is testing chatbots or agents with simulated users, but pricing, release date, and API access were not found.

Key points

Kairos 1 is built to simulate individual human behavior in chat, rather than play the assistant.
The video says the reported results come from an early training checkpoint.
Kairos 1 scores 86.11 on the user simulation index, compared with Persimmon at 77.82.
Its reported scores include 95 for information patterns and 84 for error reaction.
On User LM LM, Kairos scores 91.71, while Fable 5.1 scores 68.5.
On Sotopia hard, Kairos scores 47.6, compared with Fable’s 32.
The page reports 17 comparisons: Kairos leads in 13 and does not lead in four.
The video cautions that this measures user simulation, not assistant capability, and recommends waiting for independent tests.
The presenter suggests simulated-user testing as a possible use, but says this is their own guess.
The presenter could not find pricing, a release date, or API access on the page.

Chapters

Start	Chapter	Summary
00:00	What Kairos 1 is	The video introduces Kairos 1 as a model designed to play the user role and outlines its claim of simulating individual human behavior.
00:31	User simulation index	The presenter reviews the index and its component scores, noting Kairos 1’s lead over the compared models.
01:36	Behavioral benchmarks	The video compares Kairos 1 on user-message generation and social interaction tests, then notes close results against Gemini on some benchmarks.
02:37	Limitations and possible use	The presenter lists the four comparisons Kairos does not lead, cautions that the scores are company-reported, and discusses potential testing uses and missing access details.

Quotes

Kairos 1, a model built to act like the human user, not the assistant. — 00:00

these results come from an early training checkpoint. — 00:31

That is basically a tie. — 02:07

It does not say Kyros is a better assistant. — 02:37

That part is my guess. It is not something the page says. — 03:08

Concepts and entities

Kairos 1, Diachronic, Kevin Bank, User Sim Index, User LM, Sotopia hard, Gemini 3.1 Pro, Persimmon, Fable 5.1, Grok 4.6

Full transcript (original language)

00:00 All right. In this video, we look at Kairos 1, a model built to act like the human user, not the assistant. On its own benchmark, it scores 86, and the next one scores 78. Let’s go ahead and see what that means. So, most models are trained to be the assistant. Kairos 1 plays the other side of the chat. It plays the user. Diachronic says, “Today’s frontier models only understand population averages, and Kairos is built to copy the individual differences that make us human.” This is the launch post from Kevin Bank, he the founder of Diachronic. He calls it the world’s

00:31 first model built to simulate individual human behavior, and says it beats Persimmon, Astra, Fable, and Gemini across 13 simulation benchmarks. Let’s go to the page and check that. One thing before we look at numbers, the page says these results come from an early training checkpoint. So, this is not even the final model. First, the user sim index. It is the average of four human-likeness scores. Kairos 1 gets 86.11. Persimmon is 77.82. Grok 4.6 is 70.38. GPT-6 Astra is 58.18, and Fable

01:04 5.1 is 56.52. That is more than eight points clear of the next one. Now, the four parts: communication style, 82 against Persimmon’s 66, information patterns, 95 the highest number on the page, and error reaction, which is how a user reacts when the assistant gets it wrong, 84, and Grok 4.6 is at 52. Clarification is the closer one, 83.52 for Kairos, and 82.9 for Grok 4.6. So, that is a win, but a small win. Then,

01:36 the behavioral benchmarks. User LM LM tests if a model can write realistic user messages from an intent. Kairos gets 91.71, and Fable 5.1 gets 68.5. That is more than 23 points, the biggest gap on the whole page. Sotopia hard is about negotiation, collaboration, and conflict. Kairos gets 47.6, and the next best is Fable at 32, still under 50, so it is a hard test, but it is more than 15 points ahead. Now, Gemini 3.1 Pro is not in the overall chart. It shows up

02:07 here in the behavioral benchmarks, and the page says its numbers come from its published OD Sim report, while the rest come from the evaluation used for this comparison. And on some of these, the lead is tiny. Sim Arena math 71.52 against Gemini’s 71.5. That is basically a tie. Human Eval email 51.24 against 50.8. Now, the honest part, the page reports all 17 comparisons, and Kyros leads on 13, on four it does not. Human Eval opinion 42 against Fable’s 43. Soc

02:37 Sci 210 75 against Gemini’s 78. Human Eval chat 24 against Fable’s 27. And a line X 72 against GPT-6 Astra 77. Also, mind you, this is about simulating users. It does not say Kyros is a better assistant. These are also numbers presented by their chronic, and the page itself says to compare scores only within a chart. So, wait for independent tests. So, why would you want a model that acts like a user? My take is testing. You can try your chatbot or your agent against realistic users

03:08 before real people touch it. That part is my guess. It is not something the page says. What I could not find on the page is pricing, a release date, or API access, so I cannot tell you how to run it yet. So, in this video, we have seen Kyros won 13 wins out of 17 from an early checkpoint. Go ahead and check the page, link in the description, and let me know what you think. I’ll see you in the next one.